The AI-enabled Source Code Vulnerability Analysis Loop: SEC543 and Anthropic’s Guidance

Anthropic published a guide on May 27 for using Claude Opus (and other models) to find and fix vulnerabilities in source code, along with an open-source repo that implements it. The guide is worth reading because of how ordinary the workflow looks. It is not a demo of a model finding exotic bugs. It is a practical six-step loop for finding flaws in source code to find and eliminate vulnerabilities.

The steps: build a threat model, build a sandbox, run discovery, verify which findings are actually exploitable, triage and deduplicate, then patch and confirm the fix holds.  Discovery is now straightforward to parallelize, and the work has moved to verification, triage, and patching. As of May 22, Anthropic reported disclosing 1,596 vulnerabilities from its open-source scanning, of which 97 were patched.

That gap is the whole story. A model can generate a hundred candidate findings before lunch, and most of the effort is everything after: proving a finding is real, ruling out the ones that are not, ranking what matters, and writing a patch that fixes the root cause without breaking the service.

The skill is in the loop, not the scan

Read the guide closely and the hard parts are not the discovery step. They are the judgment steps around it. Threat modeling decides what even counts as a vulnerability in your environment, and Anthropic notes that when the threat model was well defined, one team's findings were exploitable 90 percent of the time. Verification has to be run by a separate agent that tries to disprove the finding, because a discovery agent asked to verify its own work filters out real bugs. Triage has to deduplicate by root cause and rank by reachability and blast radius, or engineers stop trusting the reports. Patching needs a failing test first, then an adversarial recheck.

Anthropic also lists penetration testing among the defensive uses of these models, simulating an attack to confirm how a vulnerability would actually be exploited. Leveraging AI to build a proof of concept on a real test bed is what moved one team's false positive rate toward zero, letting them focus on real, exploitable vulnerabilities and their associated fixes.

SANS SEC543: Practical Techniques Applied, Hands-On

The brand-new SANS SEC543: AI-Assisted Source Code Analysis and Exploitation for Penetration Testers course pioneered this analytic loop to teach cyber practitioners how to leverage AI to find flaws in source code. The class does not assume a software development background. The course covers configuring a secure AI agent stack, mapping unfamiliar code, and using AI coding agents to find logic flaws and build working exploits, all while avoiding false positives and hallucinations. Those map onto the threat model, sandbox, discovery, verification, and exploitation steps the guide describes. It is co-authored by Josh Wright, Ed Skoudis, Chris Davis, and Evan Booth at Counter Hack.

The Anthropic guide and the repo are worth working through on their own, whatever you make of the course. If you want to run the loop yourself, the harness includes a quickstart in Claude Code. Links below.

Anthropic, Using LLMs to secure source code (May 27 guide): https://claude.com/blog/using-llms-to-secure-source-code

defending-code-reference-harness (open-source repo): https://github.com/anthropics/defending-code-reference-harness

SEC543: https://www.sans.org/cyber-security-courses/ai-source-code-analysis-exploitation-pentesters